Learn Logic

Oxford and Cambridge Researchers proved every LLM trained only on AI-generated content develops an irreversible disorder. They call it "Model Collapse" Researchers took a small language model and fed it its own output for nine generations. Each new model learned only from what the one before it wrote. By generation nine the model had forgotten what it was talking about. Ask it about medieval church towers and it answers with a list of jackrabbits. They called it model collapse. The name stuck. The mechanic is simple. Every model slightly overproduces the common stuff and underproduces the rare stuff. Train the next one on that output and the rare stuff shrinks again. Do that enough times and the edges of reality disappear. The rare stuff is where the useful things live. Minority languages, unusual cases, the outliers that make a model sharp instead of average. Now look at what gets published online. Ahrefs checked 900,000 new web pages last year and found AI-generated text in 74 percent of them. Every scraper feeding the next training run picks that up. The labs know this. That's why they're paying for human archives. News Corp alone got a reported 250 million dollars from OpenAI for five years of access. The paper called it a first mover advantage. Whoever trained before the flood holds an asset nobody can recreate. The paper also found half a fix. When the researchers kept just 10 percent real human data in every generation, the damage dropped to minor. Real data isn't optional. It's the anchor. So the thing being burned right now isn't compute or money. It's the supply of human writing that hasn't been through a model. Every time someone posts machine output, the next model gets a little more average. That isn't a moral point. It's a measurement. The paper is two years old. The experiment it describes is still running, and we're the training data.